EP0733982A1 - Verfahren zur Beschleunigung der Ausführungsgeschwindigkeit Neuronalnetzwerken für korrelierte Signalverarbeitung - Google Patents

Verfahren zur Beschleunigung der Ausführungsgeschwindigkeit Neuronalnetzwerken für korrelierte Signalverarbeitung Download PDF

Info

Publication number
EP0733982A1
EP0733982A1 EP96104445A EP96104445A EP0733982A1 EP 0733982 A1 EP0733982 A1 EP 0733982A1 EP 96104445 A EP96104445 A EP 96104445A EP 96104445 A EP96104445 A EP 96104445A EP 0733982 A1 EP0733982 A1 EP 0733982A1
Authority
EP
European Patent Office
Prior art keywords
neuron
neurons
activation
quantized
quant
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
EP96104445A
Other languages
English (en)
French (fr)
Other versions
EP0733982B1 (de
Inventor
Dario Albesano
Roberto Gemello
Franco Mana
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Telecom Italia SpA
Original Assignee
CSELT Centro Studi e Laboratori Telecomunicazioni SpA
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by CSELT Centro Studi e Laboratori Telecomunicazioni SpA filed Critical CSELT Centro Studi e Laboratori Telecomunicazioni SpA
Publication of EP0733982A1 publication Critical patent/EP0733982A1/de
Application granted granted Critical
Publication of EP0733982B1 publication Critical patent/EP0733982B1/de
Anticipated expiration legal-status Critical
Expired - Lifetime legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means

Definitions

  • the present invention relates to automatic signal recognition systems and in particular it concerns a method of speeding up the execution of neural networks to process correlated signals.
  • a neural network is a parallel processing model reproducing, in extremely simplified form, the cerebral cortex organization.
  • a neural network comprises multiple processing units, known as neurons, strongly interconnected by means of links of various intensities, called synapses or interconnection weights.
  • Neurons in general are arranged along a multi-level structure, with an input level, one or more intermediate levels and an output level. Starting from the input units, which receive the signal to be processed, processing propagates to the subsequent levels of the network up to the output units, which provide the result.
  • a neural network is not programmed, but it is trained by means of a series of examples of the phenomenon to be modelled. Various embodiments of neural networks are described, for instance, in the book by D. Rumelhart “Parallel Distributed Processing", Vol. 1 Foundations, MIT Press, Cambridge, Mass., 1986.
  • Neural network technology can be applied to many sectors, such as function estimation, audio and video signal processing and classification, automatic controls, forecasting and optimisation, although it still presents some problems stemming from the lack of means that are powerful enough in terms of processing power and speed. It is well known that the execution of a neural network, when it is carried out by emulation on a sequential processor, is very burdensome, especially in cases requiring networks with many thousands of weights. If the need arises to process, in real time, signals continuously varying through time, such as for example voice, video, sonar or radar signals, use of this technology takes on additional difficulties.
  • the first is aimed at decreasing the network size by pruning weights and units, as described for example by R. Reed in "Pruning Algorithms - A Survey” in IEEE Transactions on Neural Networks, Vol. 4, no. 5, 1993. These methods, however, have limited effectiveness since the number of weights and units that can be pruned without incurring a degradation in performance is often very limited.
  • a second route is based on implementation of the neural network on a VLSI chip, by exploiting its intrinsic potential for parallelisation. This method is, potentially, very promising, but it is not very mature yet. It also entails the use of specialised hardware, which is often very expensive and not easy to integrate with commercial processors.
  • a third route is the use of specialised hardware of multi-processor type, by distributing the execution of the neural network among various processors.
  • this possible solution also requires non-standard hardware, which is costly and difficult to integrate with commercial platforms like personal computers or workstations.
  • the aforesaid drawbacks are obviated by the method of speeding up the execution of neural network for correlated signal processing, according to the present invention, which allows speeding up the execution of a wide class of neural networks to process sequential input signals evolving slowly through time, such as, for instance, voice, radar, sonar, video signals, and which requires no specialised, costly or hard to find hardware.
  • the object of the present invention is to provide a method of speeding up the execution of neural networks for correlated signal processing, as defined in the characterising part of claim 1.
  • the idea the method is based upon is the following: since the input signal is sequential and evolves slowly and continuously through time, it is not necessary to compute again all the activation values of all neurons for each input, but rather it is enough to propagate through the network the differences with respect to the previous input. That is, the operation does not consider the absolute neuron activation values at time t, but the differences with respect to activation values at time t-1. Therefore at any point of the network, if a neuron has at time t an activation that is sufficiently similar (preferably identical) to that of time t-1, then the neuron will not propagate any signal forward.
  • the execution of the neural network can be speeded up by propagating significant activation differences, and this allows saving up to two thirds of execution time in case of speech recognition.
  • This method requires a very small amount of auxiliary memory and it does not entail an appreciable degradation in performance, as it was experimentally verified.
  • Figure 1 shows a Multi-layer Perception neural network like the one described in the already mentioned book by D. Rumelhart "Parallel Distributed Processing".
  • the network input is a signal sampled in time and the network output are values corresponding to the desired processing, for instance input signal classification.
  • FIG. 2 shows a single neuron i with its forward connections, along which it propagates the activation differences, and with its memory structures M1 i and M2 i required for the speeding up method.
  • M1 i contains the activation value at time t, o i (t), as in conventional neural networks, and M2 i the value at the preceding time t-1, o i (t-1).
  • the other neurons in the network also have similar memory structures, for example M1 k and M2 k for neuron k.
  • Figure 3 depicts the quantization of the set of output values (co-domain) of the sigmoid transfer function of the neuron, with the purpose of quantizing the activation levels of the neurons, thus making it possible to recognise the condition of activation similarity at times t and t-1, required in order no signal is propagated.
  • the neurons do not propagate any signal when the quantized values at times t and t-1 are identical. The elementary operation to speed up execution of the network is thus accomplished.
  • the number of quantization values must be estimated empirically: the smaller it is, the more the method accelerates; however, it cannot be excessively small to avoid a degradation in performance. In the case of realistic Multi-layer Perceptron networks, with about 50,000 - 100,000 weights, this number can vary from about 25 to about 50.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Neurology (AREA)
  • Image Analysis (AREA)
  • Feedback Control In General (AREA)
  • Complex Calculations (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)
  • Measurement Of Velocity Or Position Using Acoustic Or Ultrasonic Waves (AREA)
  • Radar Systems Or Details Thereof (AREA)
EP96104445A 1995-03-22 1996-03-20 Verfahren zur Beschleunigung der Ausführungsgeschwindigkeit von Neuronalnetzwerken für korrelierte Signalverarbeitung Expired - Lifetime EP0733982B1 (de)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
IT95TO000216A IT1280816B1 (it) 1995-03-22 1995-03-22 Metodo per velocizzare l'esecuzione di reti neurali per il trattamento di segnali correlati.
ITTO950216 1995-03-22

Publications (2)

Publication Number Publication Date
EP0733982A1 true EP0733982A1 (de) 1996-09-25
EP0733982B1 EP0733982B1 (de) 1999-06-02

Family

ID=11413390

Family Applications (1)

Application Number Title Priority Date Filing Date
EP96104445A Expired - Lifetime EP0733982B1 (de) 1995-03-22 1996-03-20 Verfahren zur Beschleunigung der Ausführungsgeschwindigkeit von Neuronalnetzwerken für korrelierte Signalverarbeitung

Country Status (6)

Country Link
US (1) US5742739A (de)
EP (1) EP0733982B1 (de)
JP (1) JPH08272759A (de)
CA (1) CA2172199C (de)
DE (2) DE733982T1 (de)
IT (1) IT1280816B1 (de)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2003073416A1 (en) * 2002-02-28 2003-09-04 Loquendo S.P.A. Method for accelerating the execution of speech recognition neural networks and the related speech recognition device
WO2004057573A1 (en) * 2002-12-23 2004-07-08 Loquendo S.P.A. Method of optimising the execution of a neural network in a speech recognition system through conditionally skipping a variable number of frames
CN105306779A (zh) * 2015-10-27 2016-02-03 西安电子科技大学 基于压缩感知和索引置乱的图像加密方法

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6516309B1 (en) * 1998-07-17 2003-02-04 Advanced Research & Technology Institute Method and apparatus for evolving a neural network
US9627532B2 (en) * 2014-06-18 2017-04-18 Nuance Communications, Inc. Methods and apparatus for training an artificial neural network for use in speech recognition
WO2017149722A1 (ja) * 2016-03-03 2017-09-08 三菱電機株式会社 演算装置および演算方法
US10949737B2 (en) * 2016-07-13 2021-03-16 Samsung Electronics Co., Ltd. Method for neural network and apparatus performing same method
DE102017206892A1 (de) * 2017-03-01 2018-09-06 Robert Bosch Gmbh Neuronalnetzsystem
US11030518B2 (en) * 2018-06-13 2021-06-08 United States Of America As Represented By The Secretary Of The Navy Asynchronous artificial neural network architecture
WO2020240687A1 (ja) * 2019-05-28 2020-12-03 株式会社ソシオネクスト 演算処理方法、演算処理装置及びプログラム

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH02207369A (ja) * 1989-02-08 1990-08-17 Hitachi Ltd ニューラルネットワーク計算機
US5313558A (en) * 1991-08-13 1994-05-17 Adams James L System for spatial and temporal pattern learning and recognition
DE4224621C2 (de) * 1992-07-25 1994-05-05 Boehringer Mannheim Gmbh Verfahren zur Analyse eines Bestandteils einer medizinischen Probe mittels eines automatischen Analysegerätes
US5559811A (en) * 1994-09-14 1996-09-24 Lucent Technologies Inc. Method for identifying untestable and redundant faults in sequential logic circuits.

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
GHORBANI A A ET AL: "INCREMENTAL COMMUNICATION FOR MULTILAYER NEURAL NETWORKS", IEEE TRANSACTIONS ON NEURAL NETWORKS, vol. 6, no. 6, 1 November 1995 (1995-11-01), pages 1375 - 1385, XP000536108 *
GHORBANI A A ET AL: "Training artificial neural networks using variable precision incremental communication", PROCEEDINGS OF 1994 IEEE INTERNATIONAL CONFERENCE ON NEURAL NETWORKS (ICNN'94), ORLANDO, FL, USA, 27 JUNE-29 JUNE 1994, ISBN 0-7803-1901-X, 1994, NEW YORK, NY, USA, IEEE, USA, pages 1409 - 1414 vol.3, XP002007127 *
KWAN H K ET AL: "MULTIPLIERLESS MULTILAYER FEEDFORWARD NEURAL NETWORKS", PROCEEDINGS OF THE MIDWEST SYMPOSIUM ON CIRCUITS AND SYSTEMS, DETROIT, AUG. 16 - 18, 1993, vol. VOL. 2, no. SYMP. 36, 16 August 1993 (1993-08-16), INSTITUTE OF ELECTRICAL AND ELECTRONICS ENGINEERS, pages 1085 - 1088, XP000499764 *
MURRAY A F: "SILICON IMPLEMENTATIONS OF NEURAL NETWORKS", IEE PROCEEDINGS F. COMMUNICATIONS, RADAR & SIGNAL PROCESSING, vol. 138, no. 1 PART F, 1 January 1991 (1991-01-01), pages 3 - 12, XP000179653 *
MYUNG WON KIM ET AL: "An asynchronous inter-processor communication based, input recycling parallel architecture for large scale neural network simulation", PROCEEDINGS OF WORLD CONGRESS ON NEURAL NETWORKS, SAN DIEGO, CA, USA, 5-9 JUNE 1994, ISBN 0-8058-1745-X, 1994, HILLSDALE, NJ, USA, LAWRENCE ERLBAUM ASSOCIATES, USA, pages II/576 - 83 vol.2, XP000574437 *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2003073416A1 (en) * 2002-02-28 2003-09-04 Loquendo S.P.A. Method for accelerating the execution of speech recognition neural networks and the related speech recognition device
US7827031B2 (en) 2002-02-28 2010-11-02 Loquendo S.P.A. Method for accelerating the execution of speech recognition neural networks and the related speech recognition device
WO2004057573A1 (en) * 2002-12-23 2004-07-08 Loquendo S.P.A. Method of optimising the execution of a neural network in a speech recognition system through conditionally skipping a variable number of frames
US7769580B2 (en) 2002-12-23 2010-08-03 Loquendo S.P.A. Method of optimising the execution of a neural network in a speech recognition system through conditionally skipping a variable number of frames
CN105306779A (zh) * 2015-10-27 2016-02-03 西安电子科技大学 基于压缩感知和索引置乱的图像加密方法

Also Published As

Publication number Publication date
DE69602662T2 (de) 1999-11-18
IT1280816B1 (it) 1998-02-11
ITTO950216A1 (it) 1996-09-22
JPH08272759A (ja) 1996-10-18
DE69602662D1 (de) 1999-07-08
CA2172199C (en) 1999-07-20
EP0733982B1 (de) 1999-06-02
ITTO950216A0 (it) 1995-03-22
DE733982T1 (de) 1997-03-13
CA2172199A1 (en) 1996-09-23
US5742739A (en) 1998-04-21

Similar Documents

Publication Publication Date Title
US5095443A (en) Plural neural network system having a successive approximation learning method
EP0314170B1 (de) Mehrschichtiges Neuronalnetzwerk mit dynamischer Programmierung
US5479571A (en) Neural node network and model, and method of teaching same
EP0733982A1 (de) Verfahren zur Beschleunigung der Ausführungsgeschwindigkeit Neuronalnetzwerken für korrelierte Signalverarbeitung
EP0333798B1 (de) Einrichtung zur mustererkennung
US20040059695A1 (en) Neural network and method of training
Bodenhausen et al. The tempo 2 algorithm: Adjusting time-delays by supervised learning
US5528700A (en) Character recognition system based on a neural network
US5613042A (en) Chaotic recurrent neural network and learning method therefor
US5745874A (en) Preprocessor for automatic speech recognition system
US5157275A (en) Circuit employing logical gates for calculating activation function derivatives on stochastically-encoded signals
Köhn et al. Employing the Z-transform to optimize the calculation of the synaptic conductance of NMDA and other synaptic channels in network simulations
JP2736361B2 (ja) ニューラルネット構成方法
JP3328953B2 (ja) 非線形シナプスニューロン、そのデバイス及びそれを用いた予測方法
Bodenhausen et al. Learning the architecture of neural networks for speech recognition.
Omori et al. Noise effect on memory recall in dynamical neural network model of hippocampus
Toulson et al. A Trading System for FTSE-100 Futures Using Neural Networks and Wavelets
PIRAMUTHU et al. Improving connectionist learning with symbolic feature construction
USH1415H (en) Signal processor/analyzer with a neural network coupled to an acoustic charge transport (act) device (act)
RU1815658C (ru) Устройство дл морфодинамического моделировани нейрона
Kohara Selective presentation learning for forecasting by neural networks
Drago et al. On the convergence of a growing topology neural algorithm
JP3343626B2 (ja) ファジィ推論のためのニューラルネットワーク
JPH08123885A (ja) 神経細胞機能素子
US5963930A (en) Apparatus and method for enhancing transfer function non-linearities in pulse frequency encoded neurons

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): DE FR GB IT NL SE

17P Request for examination filed

Effective date: 19961008

EL Fr: translation of claims filed
TCNL Nl: translation of patent claims filed
DET De: translation of patent claims
GRAG Despatch of communication of intention to grant

Free format text: ORIGINAL CODE: EPIDOS AGRA

GRAG Despatch of communication of intention to grant

Free format text: ORIGINAL CODE: EPIDOS AGRA

GRAH Despatch of communication of intention to grant a patent

Free format text: ORIGINAL CODE: EPIDOS IGRA

GRAH Despatch of communication of intention to grant a patent

Free format text: ORIGINAL CODE: EPIDOS IGRA

17Q First examination report despatched

Effective date: 19981020

GRAA (expected) grant

Free format text: ORIGINAL CODE: 0009210

ITF It: translation for a ep patent filed
AK Designated contracting states

Kind code of ref document: B1

Designated state(s): DE FR GB IT NL SE

REF Corresponds to:

Ref document number: 69602662

Country of ref document: DE

Date of ref document: 19990708

ET Fr: translation filed
PLBE No opposition filed within time limit

Free format text: ORIGINAL CODE: 0009261

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: NO OPPOSITION FILED WITHIN TIME LIMIT

26N No opposition filed
REG Reference to a national code

Ref country code: GB

Ref legal event code: IF02

REG Reference to a national code

Ref country code: FR

Ref legal event code: PLFP

Year of fee payment: 20

PGFP Annual fee paid to national office [announced via postgrant information from national office to epo]

Ref country code: DE

Payment date: 20150317

Year of fee payment: 20

Ref country code: NL

Payment date: 20150309

Year of fee payment: 20

Ref country code: IT

Payment date: 20150226

Year of fee payment: 20

PGFP Annual fee paid to national office [announced via postgrant information from national office to epo]

Ref country code: GB

Payment date: 20150318

Year of fee payment: 20

Ref country code: FR

Payment date: 20150309

Year of fee payment: 20

Ref country code: SE

Payment date: 20150311

Year of fee payment: 20

REG Reference to a national code

Ref country code: DE

Ref legal event code: R071

Ref document number: 69602662

Country of ref document: DE

REG Reference to a national code

Ref country code: NL

Ref legal event code: MK

Effective date: 20160319

REG Reference to a national code

Ref country code: GB

Ref legal event code: PE20

Expiry date: 20160319

PG25 Lapsed in a contracting state [announced via postgrant information from national office to epo]

Ref country code: GB

Free format text: LAPSE BECAUSE OF EXPIRATION OF PROTECTION

Effective date: 20160319

REG Reference to a national code

Ref country code: SE

Ref legal event code: EUG